Abstract
A unified ranking scheme for effective video retrieval is proposed, in which low-level visual feature terms and high-level image category features are combined organically to inspire effective retrieval in the manner of semantics. By taking these features as a joint fact of document relevance, the BM25 model, popular in text retrieval, is employed to determine a mixed similarity rank of video documents. Experiments using the well-known TRECVID retrieval dataset have validated the superiority of the methodology.